Abstract
For a caching system with multiple users, we aim to characterize the memory-rate tradeoff for caching with uncoded cache placement, under nonuniform file popularity. Focusing on the modified coded caching scheme (MCCS) recently proposed by Yu, etal., we formulate the cache placement optimization problem for the MCCS to minimize the average delivery rate under nonuniform file popularity, restricting to a class of popularity-first placements. We then present two information-theoretic lower bounds on the average rate for caching with uncoded placement, one for general cache placements and the other restricted to the popularity-first placements. By comparing the average rate of the optimized MCCS with the lower bounds, we prove that the optimized MCCS attains the general lower bound for the two-user case, providing the exact memory-rate tradeoff. Furthermore, it attains the popularity-first-based lower bound for the case of general K users with distinct file requests. In these two cases, our results also reveal that the popularity-first placement is optimal for the MCCS, and zero-padding used in coded delivery incurs no loss of optimality. For the case of K users with redundant file requests, our analysis shows that there may exist a gap between the optimized MCCS and the lower bounds due to zero-padding. We next fully characterize the optimal popularity-first cache placement for the MCCS, which is shown to possess a simple file-grouping structure and can be computed via an efficient algorithm using closed-form expressions. Finally, we extend our study to accommodate both nonuniform file popularity and sizes, where we show that the optimized MCCS attains the lower bound for the two-user case, providing the exact memory-rate tradeoff. Numerical results show that, for general settings, the gap between the optimized MCCS and the lower bound only exists in limited cases and is very small.
Introduction
In modern content delivery networks, caching is a vital tool to alleviate network congestion. By storing parts of popular files near the end-user during off-peak hours, servers can reduce the "delivery rate" (the amount of data sent over a shared link) during peak times. While early research into Coded Caching (CC) assumed all files were equally popular, real-world traffic is dominated by a few viral videos, following a nonuniform distribution. This paper addresses the optimization of the Modified Coded Caching Scheme (MCCS) under these realistic, nonuniform demands.
System Model
The system consists of a central server storing $N$ files, which are delivered to $K$ users via a shared bottleneck link. Each user has a local cache of size $M$. The core challenge lies in the "placement phase": deciding which parts of which files to store in each user's cache to minimize the average delivery rate when file popularities are skewed.
Methodology
The authors approach the problem through a combination of optimization and information theory:
- Optimization Framework: The cache placement problem is formulated as a Linear Program (LP) to minimize the average delivery rate. The search is restricted to "popularity-first" placements—where more popular files are allocated more cache space.
- Lower Bounds: To benchmark performance, the authors derive two new information-theoretic lower bounds on the average delivery rate for any system using uncoded placement.
- Structural Analysis: They prove that the optimal placement strategy follows a three-file-group structure. Regardless of the total number of files, the optimal strategy categorizes them into "most popular," "moderately popular," and "non-popular" groups.
- Efficient Algorithm: Instead of relying on expensive solvers, the authors derive closed-form expressions for these groups, reducing the optimization to a simple one-dimensional search.
Key Insights and Results
1. The Optimality of MCCS
The authors prove that for a two-user system, MCCS achieves the absolute lower bound, providing the exact memory-rate tradeoff. This validates MCCS as a theoretically sound foundation for asymmetric caching systems.
2. Zero-Padding and Redundancy
A major discovery is that zero-padding—padding shorter subfiles to match longer ones for XORing—incurs no loss of optimality in many regions, specifically when all users request distinct files. However, a small gap between MCCS and the lower bounds may emerge when multiple users request the same file (redundant requests).
3. Popularity-First Placement
The research provides the first analytical proof that popularity-first placement is indeed optimal for MCCS in key scenarios. This confirms a long-held heuristic in the field: prioritizing cache space for the most frequently requested content is mathematically optimal for minimizing delivery rates.
4. Performance and Scalability
Numerical results show that the optimized MCCS consistently outperforms the original Coded Caching Scheme (CCS) and other heuristic strategies like RLFU-GCC, particularly when cache sizes are small. Furthermore, the proposed algorithm can compute the optimal placement in polynomial time, a massive improvement over the exponential complexity required to evaluate MCCS directly.
Conclusion
This work provides a comprehensive characterization of the memory-rate tradeoff for MCCS under nonuniform demands. By identifying the three-group structure of optimal placement and proving the scheme's optimality in key cases, the authors offer both theoretical insights and practical tools for network engineers. Future work may explore coded placement strategies and further refine delivery for massive user bases with high request redundancy.
Summary updated